acceptodds
Under review as a conference paper at ICLR 2027

RESPONSE CALIBRATION FOR NEURAL POPULATION MODELS

Abstract

Generative models of neural dynamics predict activity well, but whether their learned operators recover the directed effects of perturbations is a separate question. We present PC-Flow, a perturbation-conditioned flow-matching model of population activity whose response operator is calibrated to randomized single-cell interventions. The operator is a finite-dose, stimulated-minus-blank mean contrast, equivalently a dose-averaged input Jacobian of the integrated flow. It estimates directed effective influence through recurrent and hidden pathways, rather than individual synaptic weights. In controlled recurrent circuits, calibration reduces response NRMSE from \CFMResponseError to \PCResponseError, compared with \GaussianCalResponseError for a calibrated Gaussian forecaster and \ResponseNetError for a shared response-only network. Gains persist across the tested recurrence and observation conditions; a mechanism-matched recurrent model remains stronger when the full state is known. Matched objective controls locate the gain in response supervision, while joint predictive scores test the additional distributional information. PC-Flow also predicts held-out simultaneous perturbations: at eight sources, its full-response error is \NonlinearPCError, versus \NonlinearGaussianError for calibrated Gaussian and \NonlinearOracleError for exact-column superposition. Direct interaction-residual tests distinguish nonlinear prediction from improved single-source estimates. These results support response-calibrated generative models as estimators of directed perturbation effects; biological validation remains open.

Primary area:
applications to neuroscience & cognitive science
Keywords:
flow matching, effective connectivity, neural dynamics

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_WU7hOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The authors propose a response calibration loss function and a training scheme suitable for learning the effects of neural perturbations. The scheme is evaluated empirically across several model families, including conditional flow matching. The evaluation is performed almost exclusively on synthetic data, including both idealized repeated-trial setups and a realistic continuous data acquisition strategy. The synthetic results confirm the utility of the modified loss function.

Critical strengths
  1. Comprehensive set of evaluations including ablations.

    The authors evaluate many distinct models, including ones that predict the effect of perturbation directly (ResponseNet, SVD shrinkage), statistical and generative models that forecast the activity of each neuron (GRU, Gaussian Transformer, Flow Matching), and mechanistic models matched to the data generator (Rate RNN). The training comparisons are also thorough, testing models with and without the response calibration term and across multiple objective ablations.

  2. Realistic data acquisition setup.

    Beyond the primary repeated-snapshot benchmark, the authors also test a training data acquisition strategy that closely matches a realistic biological protocol (Section 4.3 / Appendix B.2). In this setup, synthetic data is recorded continuously without state resets, using randomly interleaved stimulated and blank trials.

Critical weaknesses
  1. The text is very dense and hard to parse.

    It is difficult to follow the paper for a reader without direct expertise in this specific niche. The explanations in the main text are overly compressed and often not self-contained, forcing the reader to constantly jump to the appendices to understand how key baselines (such as ResponseNet) or experimental protocols actually work.

  2. No positive confirmation on real biological data.

    The evaluation providing positive evidence for the method is entirely synthetic, which only gives indirect evidence that the approach will work in practice. The authors report a negative result on a real biological recording (Appendix C), where calibration provides no improvement. The test only evaluates a linear ridge predictor rather than PC-Flow, and the lack of improvement is attributed to an unsuitable experimental protocol (missing matched blank trials) and limited recording coverage.

Questions for authors
  1. The performance comparison indicates that PC-Flow is computationally demanding at both training and inference stages. While the timings are reported, the practical trade-off for real workflows is not fully clear. Do the authors expect this scaling to be a bottleneck in realistic setups? When the extra cost of PC-Flow justified over the much faster calibrated Gaussian Transformer or ResponseNet?

  2. Did the authors try to locate other public biological datasets where the proposed method (and specifically PC-Flow) can be evaluated? Can the authors share their general thoughts on what experimental changes are needed to make their approach applicable to real recordings?

  3. Conditioning the flow on the perturbation (u=δeiu = \delta e_i vs. u=0u = 0) looks structurally very similar to Classifier-Free Guidance (CFG). Did the authors consider applying CFG at inference time to the already trained uncalibrated CFM model? Specifically, one could integrate the guided velocity field v(…∣do(0))+w(v(…∣do(δei))−v(…∣do(0)))v(\dots|\text{do}(0)) + w\left(v(\dots|\text{do}(\delta e_i)) - v(\dots|\text{do}(0))\right) with w>1w > 1 to steer the flow toward the perturbation (see https://sander.ai/2022/05/26/guidance.html). This simple experiment might clarify whether uncalibrated CFM merely underestimates the magnitude of the perturbation effect or learns an incorrect directional response pattern across neurons altogether.

  4. Why does PC-Flow outperform ResponseNet on the mean contrast? Both PC-Flow and ResponseNet use the exact same Transformer backbone and supervise the same calibration target, yet PC-Flow achieves a significantly lower response NRMSE (0.120 vs. 0.183 in Table 2). Why does learning the full generative trajectory distribution improve the estimation of the mean contrast compared to predicting the contrast directly?

  5. As shown in Table 2, response calibration dramatically improves the response NRMSE (from 0.317 down to 0.120 for PC-Flow) while leaving the forecast CRPS score virtually unchanged (0.128 vs. 0.128). Could the authors clarify why general forecasting metrics are so insensitive to perturbation calibration?

7 comments
AuthorsOpenReview ↗

Response to Reviewer WU7h (1/5)

We thank the reviewer for the positive assessment and for questions that each pointed to a real gap; the guidance and CRPS questions led to two new analyses that we think clarify the paper. Section, figure and table numbers refer to the revised PDF.

W1. Dense, not self-contained text.

We rewrote the paper. Section 3 now defines every baseline in the main text: the response network uses the same attention backbone with a deterministic head that predicts the stimulated-minus-blank response directly from history, stimulated source and dose; the Gaussian transformer uses the same backbone with a mean head and a low-rank-plus-diagonal covariance head; the rate RNN has the simulator's tanh state equation with learned weights and a learned initial-state encoder. Each experimental subsection states the protocol facts needed to read its table, a roadmap table at the start of Section 4 maps each question to its evidence and answer, Fig. 1(d) shows the model and its training.

W2 / Q2. Real data, other datasets, and required experimental changes.

Dataset search. The method needs trial-resolved recordings in which stimulated and blank trials are interleaved at random with known probabilities. Two public datasets meet this requirement, and we analyzed both (results in part 2/5):

  • Dalgleish et al. [1]: layer 2/3 of mouse barrel cortex, 11 sessions from 6 mice, pseudo-random trial types with 15% catch (no-stimulation) trials and ensembles of 5–200 targets.
  • Daie et al. [2]: mouse anterior lateral motor cortex, 84 sessions, photostimulation groups of fewer than ten neurons plus non-stimulated trials.

Other candidates lack randomized blank trials or trial-resolved data: we found no public release of the recordings in [3]; [4] releases data for reproducing figures; the CA1 data of [5] compare stimulation sites rather than stimulated versus blank trials; and public in vitro multi-electrode array datasets record spontaneous activity or pharmacological, electrical or closed-loop stimulation [6–8]. Table 24 summarizes the search.

References

[1] H. W. P. Dalgleish, L. E. Russell, A. M. Packer, A. Roth, O. M. Gauld, F. Greenstreet, E. J. Thompson, M. Häusser. How many neurons are sufficient for perception of cortical activity? eLife 9:e58889, 2020. doi:10.7554/eLife.58889. Data: doi:10.6084/m9.figshare.13128950, doi:10.6084/m9.figshare.13135505

[2] K. Daie, K. Svoboda, S. Druckmann. Targeted photostimulation uncovers circuit motifs supporting short-term memory. Nature Neuroscience 24(2):259–265, 2021. doi:10.1038/s41593-020-00776-3. Data: doi:10.25378/janelia.13546157

[3] A. Wagenmaker, L. Mi, M. Rozsa, M. S. Bull, K. Svoboda, K. Daie, M. D. Golub, K. Jamieson. Active learning of neural population dynamics using two-photon holographic optogenetics. Advances in Neural Information Processing Systems 37:31659–31687, 2024. doi:10.52202/079017-0994

[4] I. A. Oldenburg, W. D. Hendricks, G. Handy, K. Shamardani, H. A. Bounds, B. Doiron, H. Adesnik. The logic of recurrent circuits in the primary visual cortex. Nature Neuroscience 27(1):137–147, 2024. doi:10.1038/s41593-023-01510-5

[5] M. A. Triplett, E. Bäumler, A. Prodan, R. Stonis, D. S. Peterka, M. Häusser, L. Paninski. Computational optimization of two-photon holographic stimulation sites in vivo. Journal of Neural Engineering 23(2):026015, 2026. doi:10.1088/1741-2552/ae4925

[6] Y. Bar El, S. Kanner, A. Barzilai, Y. Hanein. Calcium imaging, MEA recordings, and immunostaining images dataset of neuron-astrocyte networks in culture under the effect of norepinephrine. GigaScience 8(2):giy161, 2019. doi:10.1093/gigascience/giy161

[7] B. J. Kagan, A. C. Kitchen, N. T. Tran, F. Habibollahi, M. Khajehnejad, B. J. Parker, A. Bhat, B. Rollo, A. Razi, K. J. Friston. In vitro neurons learn and exhibit sentience when embodied in a simulated game-world. Neuron 110(23):3952–3969.e8, 2022. doi:10.1016/j.neuron.2022.09.001

[8] F. E. Kapucu, A. Vinogradov, T. Hyvärinen, L. Ylä-Outinen, S. Narkilahti. Comparative microelectrode array dataset of the functional development of hPSC-derived and rat neuronal networks, version 1.0 [dataset]. G-Node, 2021. doi:10.12751/g-node.wvr3jf

AuthorsOpenReview ↗

Response to Reviewer WU7h (2/5)

W2 / Q2 (continued). [new] Results on the two datasets. We applied the paper's no-reset protocol. The calibration target is the inverse-probability-weighted contrast of each stimulated condition against blank trials [3]. Models are scaled to the data (two attention blocks, width 64). Because sessions contain only a few hundred trials, we used five-fold cross-fitting [4]: each fold's model is fitted on the other four folds and scored only against contrasts computed from the held-out fold, so predictions and references never share trials. The floor is the error that the true contrast itself would obtain against these noisy held-out references. As in the paper, the score excludes the targeted cells' own responses (Section 4.8, Table 4, Appendix C).

Response NRMSE, non-targeted cellsBarrel cortex [1]Motor cortex [2]
Pooled contrast0.9210.948
Gaussian transformer + cal.0.8820.906
CFM0.9690.962
PC-Flow0.871 [0.808, 0.938]0.897 [0.865, 0.931]
Floor0.7810.846

Calibration improves both model families (barrel cortex: PC-Flow − CFM = −0.098 [−0.126, −0.069]; calibrated − uncalibrated Gaussian = −0.069 [−0.100, −0.046]), and the ordering matches the simulations. PC-Flow and the calibrated Gaussian are statistically indistinguishable at a few hundred trials per session (−0.011 [−0.040, 0.018]). Non-targeted responses are weak and mostly suppressive in barrel cortex [1], so all models lie near the floor; the condition stimulating all 200 targets, held out of training, separates them most clearly (floor 0.624, PC-Flow 0.762, calibrated Gaussian 0.776, CFM 0.903). The revised abstract states the benefit of calibration on real data and that PC-Flow's further benefits are so far shown only in simulation.

[new] C. elegans with PC-Flow. The reviewer noted that the submitted C. elegans test [5] used only ridge regression. Under the same date-held-out protocol we now also fitted small PC-Flow, calibrated Gaussian, CFM and response-network models (group RMSE 0.10163 for PC-Flow, against 0.10175 for predicting zero change). All stay within the six-date interval of predicting zero change, as expected for recordings without blank trials (Appendix C.4).

What experiments need. We added these requirements to the Discussion, in order of importance: (1) randomly interleaved blank trials with logged assignment probabilities, because the target is a randomized contrast rather than a pre/post difference; (2) trial selection and grouping based only on pre-stimulus information; (3) a fixed pulse duration and power, ideally with a few doses; (4) repeated stimulation of the same targets; (5) sampling faster than the response, with stimulation artifacts shorter than the response peak; (6) random subsets of a multi-target set for interaction tests; (7) stable cell identities across the session. The barrel-cortex protocol [1] meets (1)–(4) and most of (5).

References

[1] H. W. P. Dalgleish, L. E. Russell, A. M. Packer, A. Roth, O. M. Gauld, F. Greenstreet, E. J. Thompson, M. Häusser. How many neurons are sufficient for perception of cortical activity? eLife 9:e58889, 2020. doi:10.7554/eLife.58889. Data: doi:10.6084/m9.figshare.13128950, doi:10.6084/m9.figshare.13135505

[2] K. Daie, K. Svoboda, S. Druckmann. Targeted photostimulation uncovers circuit motifs supporting short-term memory. Nature Neuroscience 24(2):259–265, 2021. doi:10.1038/s41593-020-00776-3. Data: doi:10.25378/janelia.13546157

[3] A. Boruvka, D. Almirall, K. Witkiewitz, S. A. Murphy. Assessing time-varying causal effect moderation in mobile health. Journal of the American Statistical Association 113(523):1112–1121, 2018. doi:10.1080/01621459.2017.1305274

[4] V. Chernozhukov, D. Chetverikov, M. Demirer, E. Duflo, C. Hansen, W. Newey, J. Robins. Double/debiased machine learning for treatment and structural parameters. The Econometrics Journal 21(1):C1–C68, 2018. doi:10.1111/ectj.12097

[5] F. Randi, A. K. Sharma, S. Dvali, A. M. Leifer. Neural signal propagation atlas of Caenorhabditis elegans. Nature 623(7986):406–414, 2023. doi:10.1038/s41586-023-06683-4. Data: doi:10.17605/OSF.IO/E2SYT

AuthorsOpenReview ↗

Response to Reviewer WU7h (3/5)

Q1. Is PC-Flow's cost a bottleneck, and when is it justified?

Where the cost is. On the profiled CPU (Table 5), a PC-Flow training update costs 18.9 s on average, against 0.91 s for the calibrated Gaussian and 0.86 s for the response network, because every fourth update integrates 16 Heun steps (a second-order ODE solver) for eight pulse/blank sample pairs. Reading out one history–source contrast with 32 pairs takes 21.8 s, against 0.027 s for the Gaussian's analytic mean and 0.014 s for the response network. The readout is the main bottleneck: 48 histories × 20 sources × 256 pairs amount to tens of CPU hours per circuit if run serially. Neuron attention also grows quadratically with the number of recorded cells. The first remedies are batching sources and sharing blank trajectories across sources, which removes about half of the trajectories; the revised Discussion states this.

[new] Reducing the cost. PC-Flow-lite, which calibrates on every 8th instead of every 4th update with 4 instead of 8 pairs and 8 instead of 16 integration steps, costs about 2.98 s per update, three times the calibrated Gaussian, and reaches response error 0.133 [0.121, 0.147], versus 0.120 for the default and 0.190 for the calibrated Gaussian; giving the calibrated Gaussian PC-Flow's training time leaves it at 0.188 (Fig. 6d). Reading out with 32 or 8 pairs instead of 256 cuts readout time 8× or 32× and gives 0.129 or 0.156 (Figure 7(d)), both still below the calibrated Gaussian.

When it is justified. For population-average single-source responses with ample repeats, the calibrated Gaussian or the response network is the better choice: their readout is orders of magnitude cheaper and, on the real data in part 2/5, equally accurate. PC-Flow is worth its cost when a question requires:

  • state-dependent responses: the error on the state-dependent part of the response (each model's mean over states removed) is 0.127 for PC-Flow versus 0.203 for the calibrated Gaussian without resets;
  • joint or tail-event predictions: at dose 4, PC-Flow's advantage in energy score, a proper score for joint predictive distributions [1], is about six times larger than at dose 0.4, and its Brier score for tail events is 0.071 versus 0.096;
  • few repeats or partial observation, where the absolute gap is largest (Fig. 3b, Fig. 6a);
  • predictions for combined pulses, where its error on the nonadditive part of eight-source responses is 0.058 versus 0.150.

References

[1] T. Gneiting, A. E. Raftery. Strictly proper scoring rules, prediction, and estimation. Journal of the American Statistical Association 102(477):359–378, 2007. doi:10.1198/016214506000001437

Reviewer_MzGFOpenReview ↗
Rating
4(+2)
Confidence
2
Informativeness
3
Presentation
2(+1)
Soundness
3
Read review
Paper summary

The paper proposes a flow-matching model of neural population activity (PC-Flow), with the target of predicting the response of a recurrent circuit to stimulation of single neurons, especially effects that propagate through indirect and unobserved pathways. The model takes as input a short, noisy history of the observed neurons, together with the stimulus, and provides a prediction of the activity at later times. The main addition to standard flow matching is a calibration loss, which matches the model's stimulated-minus-blank response to the response measured in randomized pulse and blank trials. Using simulated excitatory-inhibitory rate circuits, the authors compared their model to many other models, mainly the same flow model without calibration (CFM), calibrated Gaussian and GRU forecasters, a network that learns the response directly, and a mechanistic tanh rate RNN with and without calibration. The comparison covers the error in the predicted response averaged over test histories, distributional forecast scores (CRPS and energy score), the dependence of the response on the network state, robustness to observation fraction, dose and recurrent gain, and the response to simultaneous stimulation of several neurons. The authors showed that in some key aspects, most notably the primary response error, the flow-matching model trained with the calibration loss reaches the best results.

Critical strengths

The paper brings modern generative machine learning, flow matching in particular, to the modeling of neural circuits and perturbation experiments. Connecting these tools to an experimental question in neuroscience, namely how a circuit responds to targeted stimulation, is valuable for the community.

The main idea of the method is simple and well motivated. The forecasting loss learns the natural dynamics of the circuit, while the calibration loss targets only the effect of the stimulus, defined as the difference between stimulated trials and matched blank trials. This separation makes the causal effect of the stimulation an explicit training target, rather than hoping that it emerges from good forecasting. The use of a split-half product to obtain an unbiased estimate of the calibration loss from noisy trials is a nice technical detail.

The experimental evaluation is extensive and thorough. It includes many baselines, among them a mechanistic model that matches the simulator, controls that isolate the effect of the calibration loss with the same architecture (CFM vs. PC-Flow) and shuffled source labels, robustness tests across observation fraction, dose, recurrent gain and number of repeats, and bootstrap confidence intervals over independent circuits. The authors also report results that do not favor their method, such as the advantage of the rate RNN when the full state is known and the null result on the C. elegans recordings.

Critical weaknesses

The presentation of the paper is poor. Both the abstract and the discussion are dense with detailed measurements and jargon, while the main questions and the scientific implications are absent. The main idea and mechanism, while quite intuitive, are introduced in a complicated way that makes them difficult to understand. A major rewriting of the paper, uploaded as a revision during the discussion period, would greatly improve my assessment, even without changing any of the scientific results.

Regarding the main scientific points:

  1. The main results of the paper rely on the ability to reset the network, including its unobserved neurons, to exactly the same state many times, and to repeat the stimulation from each of these states (Appendix B.1, about 70k trials per circuit). This protocol is not experimentally possible, even in principle. The paper does include a setting without resets (continuing randomized stimulation, Section 4.3 and Appendix B.2), but there the experiments have a different setup and are not compared to all competitors. The main results should present a case where the stimulus is repeated, but the state of the network is sampled from a distribution broad enough to reflect a realistic experimental setup, where a return to a homeostatic equilibrium may be possible, but a complete reset is not.

  2. The motivating claim of the paper is the prediction of directed influence through recurrent and hidden pathways. However, the main text reports the results on indirectly connected pairs only in a single sentence comparing PC-Flow with CFM (Section 4.5), while the full stratification appears only in the appendix (Table 9). These results are significantly weaker than the main results, although this is admittedly a harder task. The table also does not contain all the main competitors from the main text. If this is the main motivation, the results should be completed and discussed thoroughly in the main text.

  3. If I understand correctly, the main response error averages the predicted and true responses over the test histories before taking the Frobenius norm of their difference over source-target pairs. This is an unusual way to measure accuracy; the more natural choice would be the mean squared error at each history, with the average taken outside, so that errors at different histories cannot cancel. It is true that this measurement has a floor: the reference at each history is computed from the full network state, including the unobserved neurons, so even a perfect model would have a nonzero error. However, in simulation this floor can be estimated by sampling the unobserved state given the observed history, and then either subtracted from the measurement or stated as a baseline.

  4. The main motivation of the paper is the stimulation of biological neurons. However, a substantial comparison of the method against other methods on neural data is lacking. The only analysis of real data is a short check on C. elegans recordings (Section 4.6 and Appendix C), which uses ridge regression rather than PC-Flow and shows no improvement from calibration. This analysis should be extended along the lines of the main text, with PC-Flow and the main competitors. Since the authors note that these recordings lack matched blank trials, other datasets should also be analyzed. From a short search, photostimulation data in mouse cortex is available, for example in Wagenmaker et al. (NeurIPS 2024, motor cortex) and Oldenburg et al. (Nature Neuroscience 2024, visual cortex), and in-vitro recordings (brain in a dish) may also be available.

Questions for authors

My questions mainly address the points raised in the weaknesses:

  1. Could the authors present the setting without resets (Section 4.3) as a main result, including all main models, and evaluate the response error at individual held-out states rather than only within the three pretreatment groups?

  2. Could the authors extend Table 9 to all main models and report the errors on direct and indirect pairs separately for each horizon?

  3. Could the authors report, for all models in the main text, the mean squared error at each test history with the average taken outside, together with an estimate of its floor obtained by sampling the unobserved state given the observed history? Even if the results are similar to the current metric, I believe this is the more natural metric to report.

  4. Could the authors apply PC-Flow and the main competitors to the C. elegans recordings, and try to analyze an additional photostimulation dataset to support the claims about the method on neural data?

9 comments
AuthorsOpenReview ↗

Response to Reviewer MzGF (1/6)

We thank the reviewer for a detailed review that states exactly what would change the assessment. We have rewritten the paper as suggested and uploaded the revision. We also made the no-reset acquisition the main result for all predictors, added per-history evaluation with an estimated floor, completed the anatomical stratification, and furthermore analyzed two public photostimulation datasets with randomized blank trials.

W1 / Q1. Resets are infeasible; present the no-reset setting as the main result, with all models and individual held-out states.

We agree and have done this in the revision version. The no-reset assay (Section 4.1; Section 4.3 of the submission) already follows the protocol the reviewer describes, but our text was too compressed to show it. Each session starts at the network's equilibrium, discards 10 s, and then runs continuously while a shared input c(t), an Ornstein–Uhlenbeck process with a 500 ms time constant, keeps moving the network across states. Pulses and blanks are assigned at random at onsets separated by 500–1000 ms, with known probabilities. The state is never restored: the network relaxes toward equilibrium but never returns to the same state, and hidden neurons are never observed.

The setup differs from the repeated-trial benchmark only where removing resets requires it: recurrent gain 1.3 makes state dependence measurable, six sources give about 4000 pulses per source, and dose 0.4 raises the single-trial signal (Appendix B.2). The repeated-trial benchmark remains as a controlled benchmark (Section 4.2) that separates estimation from acquisition.

AuthorsOpenReview ↗

Response to Reviewer MzGF (2/6)

W1 / Q1 (continued). All predictors without resets. The submitted experiments produced these results for every predictor, but only three were tabulated (fitting trials, dose 0.4, six sources). Group error compares responses averaged within three pretreatment groups, the tertiles of the first principal component of the 100 ms pre-stimulus activity. State-dependent error applies the same metric after removing each model's own mean over groups, so a history-independent predictor scores exactly 1. [new] Per-history error compares responses at 48 individual held-out histories (defined under W3, part 4/6).

ModelGroup errorState-dependent errorPer-history error
Pooled contrast, SVD shrinkage, ridge VARX0.2131.0000.463 [0.417, 0.513]
Local response ridge0.1170.4850.447
Response network0.0720.2100.312
NARX0.1010.4010.401
GRU0.1100.4480.418
GRU + cal.0.0770.2480.335
Gaussian transformer0.0910.3400.388
Gaussian transformer + cal.0.0740.2030.318
Rate RNN0.0960.3570.371
Rate RNN + cal.0.0650.1400.252
CFM0.1010.4080.419
PC-Flow0.063 [0.060, 0.066]0.127 [0.108, 0.148]0.246 [0.222, 0.273]
Floor (posterior-mean oracle)——0.168 [0.151, 0.186]

The first row groups three predictors whose response does not depend on history: the pooled contrast, its low-rank SVD shrinkage, and ridge VARX (linear autoregression with stimulus input). NARX is a nonlinear autoregressive model; local response ridge regresses responses on principal components of the history. Table 2 shows the four forecaster families, and Table 10 gives all predictors with intervals.

Across fitting budgets of 2048 / 8192 / 32768 trials, PC-Flow's group error is 0.233 / 0.119 / 0.063, against 0.247 / 0.130 / 0.074 for the calibrated Gaussian and 0.240 / 0.123 / 0.065 for the calibrated rate RNN (Fig. 2a). Calibration helps every model family without resets, and the calibrated models keep their ranking from the repeated-trial benchmark.

Individual held-out states. For 48 held-out onsets per circuit, an evaluation-only simulator restores the full state, including the shared input, and runs 4096 coupled pulse/blank simulations; this reference is never used for fitting. Sampled models are read out with 2,048 pairs per history. As the reviewer proposed (details under W3), the floor comes from a particle filter that infers both hidden neurons and the shared input from the observed history. The true response varies strongly across individual states (0.35, 0.41, 0.58 and 0.89 of the mean response at 10, 30, 60 and 100 ms, versus 0.16–0.30 across the three groups on the same entries), so per-history errors exceed group errors for every model. PC-Flow and the calibrated rate RNN are the best predictors and are statistically indistinguishable (difference −0.006 [−0.034, 0.019]); both remain above the floor, which is itself well above zero because the shared input is unobserved.

AuthorsOpenReview ↗

Response to Reviewer MzGF (3/6)

W2 / Q2. Direct and indirect pairs for all main models, per horizon.

Done, from the submitted predictions (repeated-trial benchmark, 64 of 80 cells observed). Direct pairs have a synapse from the stimulated cell to the target (about 128 pairs per circuit); indirect pairs have no synapse but a directed path (about 1,132 pairs). Each stratum is normalized by its own reference norm.

ModelDirect, 10 / 30 / 60 / 100 msIndirect, 10 / 30 / 60 / 100 ms
Pooled contrast0.09 / 0.10 / 0.14 / 0.213.93 / 1.17 / 0.67 / 0.48
Response network0.05 / 0.06 / 0.09 / 0.132.02 / 0.61 / 0.38 / 0.27
GRU + cal.0.07 / 0.08 / 0.09 / 0.132.35 / 0.68 / 0.37 / 0.27
Gaussian transformer + cal.0.07 / 0.06 / 0.09 / 0.132.25 / 0.62 / 0.37 / 0.28
Rate RNN + cal.0.05 / 0.05 / 0.07 / 0.112.18 / 0.60 / 0.31 / 0.23
CFM0.15 / 0.13 / 0.16 / 0.224.22 / 1.00 / 0.57 / 0.43
PC-Flow0.04 / 0.04 / 0.05 / 0.081.53 / 0.41 / 0.23 / 0.18

Indirect relative errors are large at short lags because indirect responses start late. For lags within the 10 ms pulse, the linear expansion in Appendix A gives a direct response proportional to h2h^2 and a two-step indirect response proportional to h3h^3, so at 10 ms the indirect reference is close to zero and any error is magnified. Once indirect responses have developed (60–100 ms), PC-Flow's indirect error is 0.23 and 0.18, the lowest of all models at every horizon. Measured against the norm of the whole response matrix, the error on indirect pairs is 0.112 for PC-Flow, 0.154 for the calibrated rate RNN, 0.177 for the calibrated Gaussian and 0.288 for CFM.

[new] Hidden-mediated pairs. To test propagation through unobserved neurons directly, we selected indirect pairs for which at least half of the two-step path strength ∑k∣WjkWki∣\sum_k\lvert W_{jk}W_{ki}\rvert runs through unobserved intermediate cells kk (the three-step strength when no two-step path exists). This gives about 307 pairs per circuit with 64 cells observed and 452 with 40 observed, and none when every cell is observed. As for all indirect pairs, the 10 ms errors are omitted because these responses have barely started.

Model64 observed: 30 / 60 / 100 ms40 observed: 100 ms
Response network0.67 / 0.40 / 0.270.38
Gaussian transformer1.04 / 0.58 / 0.390.54
Gaussian transformer + cal.0.69 / 0.39 / 0.290.41
Rate RNN1.02 / 0.52 / 0.380.52
Rate RNN + cal.0.67 / 0.34 / 0.250.34
CFM1.12 / 0.61 / 0.450.62
PC-Flow0.45 / 0.25 / 0.180.26

Calibration lowers the hidden-mediated error for every model family, and PC-Flow is the most accurate at every horizon and observation level. These pairs remain harder than direct pairs. Section 4.4 now presents the full stratification in the main text (Fig. 4), as an audit of predicted responses against known anatomy rather than as connectivity recovery, and Tables 14 and 15 give every horizon, replacing the submitted Table 9.

Reviewer_XtCEOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

The paper studies whether neural population models that predict activity accurately also capture stimulation-induced responses. It proposes response calibration, which supplements predictive training with supervision of stimulus-minus-blank contrasts, and implements this approach using conditional flow matching (PC-Flow). Simulated experiments evaluate response estimation, state dependence, partial observation, and generalization to simultaneous perturbations. The results show improved response accuracy, while biological validation remains open.

Critical strengths
  1. The paper addresses an interesting and practically relevant question: whether predictive accuracy is sufficient for using learned neural models to study interventions.
  2. Calibrating experimentally defined response contrasts is a well-motivated approach that applies across model families and directly connects the learning objective to the intended intervention query.
Critical weaknesses
  1. Although PC-Flow improves response estimation in the reported simulations, the evidence for its additional distributional benefit is limited. Given the substantial computational cost, the practical benefit over simpler calibrated predictors remains insufficiently established.
  2. The main experimental section is difficult to follow. Although detailed protocols are provided in the appendix, the changing experimental settings and baseline coverage make it difficult to understand how the individual assays jointly support the central claims.
Questions for authors
  1. What motivates conditional flow matching specifically for response calibration?
  2. How does the calibrated rate RNN perform on full-response and interaction-residual errors in the strong multi-source assay?
  3. Under partial observation, how accurate are responses at individual held-out histories before averaging?
  4. Do the advantages over calibrated Gaussian and recurrent predictors persist under matched training and inference budgets?
  5. Does PC-Flow retain its advantage over calibrated Gaussian and recurrent predictors as observation noise increases and the number of repeated trials decreases?
7 comments
AuthorsOpenReview ↗

Response to Reviewer XtCE (1/4)

We thank the reviewer for all the clear and constructive questions. We have uploaded a re-written PDF (see W2). Following section, figure and table numbers refer to the revised PDF, and we will say so where we cite the submitted version. Below, response error is the normalized root-mean-square error (NRMSE) between predicted and true stimulated-minus-blank responses (Table 3); "+ cal." marks models trained with response calibration; CFM is the same flow model trained without calibration; brackets are 95% bootstrap intervals over independent circuits; [new] marks experiments added during this discussion.

W1. Limited evidence for a distributional benefit, given the cost.

We agree that the submitted evidence was limited. The submitted Table 7 (now Table 17) measured joint scores at dose 0.4, where the stimulated distribution is close to a Gaussian with the same mean and covariance (energy score 0.1638 for that moment-matched Gaussian, 0.1625 for PC-Flow). A benefit beyond the mean and covariance can only appear where the stimulus changes the shape of the distribution, so we tested a strong dose.

[new] Joint scores at dose 4. We repeated the no-reset of Section 4.1 (randomized stimulation without resetting the network state) at dose 4 instead of 0.4, with the same 12 circuits, assignment probabilities and fitting trials, refitted all models, and scored them at the same three histories per circuit. The energy score (ES) and variogram score (VS) are proper scores for the joint distribution of the 16 projected coordinates [1]. The Brier score is the squared error of a predicted probability, here for the event that a projected cell exceeds the 95th percentile of its blank distribution at 60 ms. The first row scores 256 draws from the true conditional distribution. Lower is better.

ModelESES − PC-FlowVSTail BrierES, unseen 8-source set
True distribution0.2286−0.0023 [−0.0031, −0.0016]0.02170.0510.2614
PC-Flow0.2309—0.02240.0710.2652
Gaussian transformer + cal.0.23780.0069 [0.0047, 0.0092]0.02510.0960.2741
GRU + cal.0.24020.0093 [0.0066, 0.0121]0.02540.0990.2779
CFM0.23470.0038 [0.0021, 0.0056]0.02310.0880.2716
Moment-matched Gaussian0.23490.0040 [0.0024, 0.0057]0.02490.0930.2688

The gap to the calibrated Gaussian is now about six times the dose-0.4 gap (0.0069 versus 0.0012), and PC-Flow outperforms a Gaussian given the true mean and covariance. These gain-1.3 circuits are only locally stable, so whether a strong pulse triggers a large recurrent transient depends on the unobserved state; the stimulated distribution becomes right-skewed with a heavy upper tail, which a Gaussian cannot represent (Fig. 5a,b). CFM has a similar shape but a biased mean: calibration supplies the mean, and the flow supplies the shape.

The flow also captures how the mean response depends on state and on combined inputs, which the Gaussian with the same backbone does not: the error on the state-dependent part of the response (each model's mean over states removed) is 0.127 versus 0.203 without resets (Table 2) and 0.135 versus 0.263 with known states (Table 12), and the error on the nonadditive part of eight-source responses is 0.058 versus 0.150.

Cost. Matched-budget results, a cheaper PC-Flow configuration, and guidance on when to prefer the calibrated Gaussian are in our answer to Q4 (part 4/4).

References

[1] T. Gneiting, A. E. Raftery. Strictly proper scoring rules, prediction, and estimation. Journal of the American Statistical Association 102(477):359–378, 2007. doi:10.1198/016214506000001437

AuthorsOpenReview ↗

Response to Reviewer XtCE (2/4)

W2. The experimental section is hard to follow.

We restructured Section 4 so that each subsection answers one question with its own table or figure and states the protocol facts needed to read it. We take the no-reset acquisition as the main result. A roadmap (Table 1) is below:

SectionQuestionAnswer
4.1Does calibration help when the state cannot be reset?Yes, in all four forecaster families
4.2Which design choices produce the gain?The response target, not the estimator or weights
4.3Are state-dependent responses recovered?Without resets, about 70% of their variation
4.4Are indirect and hidden-mediated responses recovered?Yes, once they have developed
4.5When does the predictive distribution add value?At strong doses
4.6What are the operating limits and the cost?Stable ranking; update cost 3–21× a calibrated Gaussian
4.7Do predictions extend to combined pulses?44% below exact superposition at 8 sources
4.8Does calibration help on real recordings?Yes; flow and Gaussian perform alike

Figures draw their models from one fixed core set of six, so the baselines no longer change from panel to panel, and all baselines (response network, Gaussian transformer, rate RNN) are now defined in Section 3, so the main text no longer depends on the appendix.

Q1. Why conditional flow matching?

Three properties matter. (i) A flow maps a noise sample to a future deterministically, so pulse and blank futures can share the same noise sample and history. Their difference then has low variance, which makes a sampled calibration loss practical with eight sample pairs per update. (ii) The stimulus enters the velocity field at every integration step and interacts with the intermediate state, so the predicted mean can depend nonlinearly on state and input. The Gaussian transformer reads its mean linearly from the same backbone features; replacing that readout by a three-layer MLP (width 256) changes its response error only from 0.190 to 0.186 [0.169, 0.204], so head capacity does not explain the gap. (iii) The flow represents the full joint distribution with 16 integration steps, which supports the joint and tail-event predictions in W1. The new Fig. 1(d) shows these steps.

Calibration itself is not specific to flows: it also improves the Gaussian transformer (0.285 to 0.190), the GRU (0.317 to 0.194) and the rate RNN (0.248 to 0.164).

Q2. Calibrated rate RNN on the strong multi-source assay.

Full is the error of the whole multi-source response; interaction is the relative error of its nonadditive part (Section 4.7).

ModelFull, k = 2 / 4 / 8Interaction, k = 4Interaction, k = 8
Rate RNN0.160 / 0.150 / 0.1410.3280.222 [0.195, 0.250]
Rate RNN + cal.0.093 / 0.087 / 0.0810.1850.129 [0.111, 0.148]
Gaussian transformer + cal.0.078 / 0.076 / 0.0780.2380.150 [0.140, 0.162]
PC-Flow0.052 / 0.048 / 0.0490.0900.058 [0.052, 0.065]
Sum of true single-source responses0.020 / 0.040 / 0.089——

Calibration reduces the rate RNN's interaction error by 42%, placing it between the calibrated Gaussian and PC-Flow. In the known-state control the rate RNN is the best model (state-dependent error 0.045 versus 0.135), but there it receives the exact initial state and is tested at dose 0.2.

We attribute the difference to extrapolation: here the model must extend weights estimated from single-source trials to dose-4 compound pulses, where small weight errors change how several inputs combine. Thus, a correct functional form helps most when the state is known and inputs stay near the linear range. Fig. 5(c,d) and Table 18 include both rate RNN rows.

AuthorsOpenReview ↗

Response to Reviewer XtCE (3/4)

Q3. Accuracy at individual held-out histories under partial observation.

In this benchmark, per-history accuracy is close to averaged accuracy when sampled models are read out with enough pairs. With 2048 readout pairs PC-Flow reaches 0.130 with 64 of 80 cells observed and 0.173 with 40, the lowest of all models; the calibrated rate RNN reaches 0.171 and 0.234. Details are shown below.

[new] Metric. We added a per-history error in which errors at different histories cannot cancel:

Ehist=14∑h∑k∥M⊙(D^kh−Dkh)∥F2∑k∥M⊙Dkh∥F2E_{\mathrm{hist}}=\frac{1}{4}\sum_{h}\sqrt{\frac{\sum_{k}\lVert M\odot(\hat D_{kh}-D_{kh})\rVert_F^2}{\sum_{k}\lVert M\odot D_{kh}\rVert_F^2}}

where DkhD_{kh} is the true response matrix (targets × sources) at test history kk and horizon h∈{10,30,60,100}h\in\lbrace 10,30,60,100\rbrace ms, D^kh\hat D_{kh} is the prediction, and the mask MM removes each stimulated cell's own response. Each per-history reference uses 4096 coupled pulse/blank simulations in order to keep the Monte Carlo error below 2%.

[new] Floor. The true response at a history depends on the full network state, but every model sees only the observed history, so even a perfect model has a nonzero error. We estimate this floor with a posterior-mean oracle: a particle filter with the true generator (20,000 particles, with proposals guided toward each next observed frame so that the weights do not collapse onto a single particle) samples the hidden state given the observed history, and the oracle averages simulated pulse-minus-blank responses over 256 posterior samples (Appendix B.5).

Result. In the repeated-trial benchmark the true response barely changes across histories: integrating the generator from the test states gives a relative variation of only 0.010–0.014 at every horizon, consistent with the near-additivity of this weak-input assay (Section 4.2). Per-history accuracy therefore matches averaged accuracy, apart from two effects the new metric exposes (64 of 80 cells observed):

ModelAveraged (Table 3)Per-history
Pooled contrast (history independent)0.3300.331 [0.302, 0.362]
Response network0.1830.195 [0.175, 0.217]
Gaussian transformer + cal.0.1900.203 [0.187, 0.220]
Rate RNN + cal.0.1640.171 [0.157, 0.186]
CFM, 256 readout pairs0.3170.340 [0.319, 0.362]
PC-Flow, 256 readout pairs0.1200.174 [0.166, 0.184]
PC-Flow, 2,048 readout pairs0.1200.130 [0.120, 0.143]
Floor—0.006 [0.005, 0.007]

First, deterministic and analytic models rise by about 0.01, which is their own variation across histories where the truth barely varies.

Second, sampled models show Monte Carlo readout noise that averaging over 48 histories hides: the readout curve in Figure 7(d) implies a per-history sampling error of 0.12 with 256 pairs, and 2048 pairs bring PC-Flow to 0.130. With 40 of 80 cells observed, the floor is 0.008, and PC-Flow scores 0.208 (0.173 with 2048 pairs), the calibrated rate RNN 0.234, the calibrated Gaussian 0.277 and CFM 0.450 (Table 13).

The metric matters most where responses depend on state. Without resets (gain 1.3, unobserved shared input), responses vary across individual histories by 0.35–0.89 of their mean, and the same evaluation gives a floor of 0.168, PC-Flow 0.246, calibrated rate RNN 0.252, calibrated Gaussian 0.318, CFM 0.419, and 0.463 for history-independent predictors (Table 2, Fig. 2c). Section 4.3 reports both metrics, the floor, and the readout budget of sampled models.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.